athena Example

This example trains an athena neural network to approximate sin(x) and logs metrics to wandb every 10 epochs.

Source: example/athena_logging/train_with_wandb.f90

Metrics logged

• epoch

• training_loss

• validation_loss

• learning_rate

Running

source tools/setup_env.sh
fpm run --example athena_logging

Key code

use athena
use wf

! initialise wandb
call wandb_init(project="wandb-fortran-athena", name="sine-demo")
call wandb_config_set("num_epochs",    num_epochs)
call wandb_config_set("num_hidden",    num_hidden)
call wandb_config_set("learning_rate", real(learning_rate, kind=8))

! training loop
do epoch = 1, num_epochs
   ! ... athena forward / backward / update ...
   if (mod(epoch, 10) == 0) then
      call wandb_log("epoch",           epoch,      step=epoch)
      call wandb_log("training_loss",   train_loss, step=epoch)
      call wandb_log("validation_loss", val_loss,   step=epoch)
      call wandb_log("learning_rate",   real(learning_rate, kind=8), step=epoch)
   end if
end do

call wandb_finish()
call wandb_shutdown()

fpm dependency

The example pulls in athena automatically via fpm:

[[example]]
name = "athena_logging"
source-dir = "example/athena_logging"
main = "train_with_wandb.f90"
[example.dependencies]
athena = { git = "https://github.com/nedtaylor/athena", tag = "v2.0.0" }